Coarse-to-Fine Visual Place Recognition

Junkun Qi, Rui Wang, Chuan Wang, Xiaochun Cao

Abstract: Visual Place Recognition (VPR) aims to locate one or more images depicting the same place in the geotagged database with a given query and is typically conducted as an image retrieval task. Currently, global-based and local-based descriptors are two mainstream representations to solve VPR. However, they still struggle against viewpoint change, confusion from similar patterns in different places, or high computation complexity. In this paper, we propose a progressive Coarse-To-Fine (CTF-VPR) framework, which has a strong ability on handling irrelevant matches and controlling time consumption. It employs global descriptors to discover visually similar references and local descriptors to filter those with similar but irrelative patterns. Besides, a region-specific representing format called regional descriptor is introduced with region augmentation and increases the possibilities of positive references with partially relevant areas via region refinement. Furthermore, during the spatial verification, we provide the Spatial Deviation Index (SDI) considering coordinate deviation to evaluate the consistency of matches. It discards exhaustive and iterative search and reduces the time consumption hundreds of times. The proposed CTF-VPR outperforms existing approaches by 2%–3% recalls on Pitts250k and Tokyo24/7 benchmarks.